Upload folder using huggingface_hub

config.json

@@ -0,0 +1,25 @@
+{
+  "architectures": [
+    "Qwen2VisionTransformerPretrainedModel"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_qwen2_vl.Qwen2VLVisionConfig",
+    "AutoModel": "modeling_qwen2_vl.Qwen2VisionTransformerPretrainedModel"
+  },
+  "depth": 32,
+  "embed_dim": 1280,
+  "hidden_act": "quick_gelu",
+  "hidden_size": 1536,
+  "in_channels": 3,
+  "in_chans": 3,
+  "initializer_range": 0.02,
+  "mlp_ratio": 4,
+  "model_type": "qwen2_vl",
+  "num_heads": 16,
+  "patch_size": 14,
+  "spatial_merge_size": 2,
+  "spatial_patch_size": 14,
+  "temporal_patch_size": 2,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.1"
+}

configuration_qwen2_vl.py

@@ -0,0 +1,57 @@
+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Qwen2VL model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class Qwen2VLVisionConfig(PretrainedConfig):
+    model_type = "qwen2_vl"
+    base_config_key = "vision_config"
+
+    def __init__(
+        self,
+        depth=32,
+        embed_dim=1280,
+        hidden_size=3584,
+        hidden_act="quick_gelu",
+        mlp_ratio=4,
+        num_heads=16,
+        in_channels=3,
+        patch_size=14,
+        spatial_merge_size=2,
+        temporal_patch_size=2,
+        initializer_range=0.02,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.depth = depth
+        self.embed_dim = embed_dim
+        self.hidden_size = hidden_size
+        self.hidden_act = hidden_act
+        self.mlp_ratio = mlp_ratio
+        self.num_heads = num_heads
+        self.in_channels = in_channels
+        self.patch_size = patch_size
+        self.spatial_merge_size = spatial_merge_size
+        self.temporal_patch_size = temporal_patch_size
+        self.initializer_range = initializer_range
+
+

image_processing_qwen2_vl.py

@@ -0,0 +1,494 @@
+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Image processor class for Qwen2-VL."""
+
+import math
+from typing import Dict, List, Optional, Union
+
+import numpy as np
+
+from ...image_processing_utils import BaseImageProcessor, BatchFeature
+from ...image_transforms import (
+    convert_to_rgb,
+    resize,
+    to_channel_dimension_format,
+)
+from ...image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    get_image_size,
+    infer_channel_dimension_format,
+    is_scaled_image,
+    make_flat_list_of_images,
+    make_list_of_images,
+    to_numpy_array,
+    valid_images,
+    validate_preprocess_arguments,
+)
+from ...utils import TensorType, logging
+from ...video_utils import VideoInput, make_batched_videos
+
+
+logger = logging.get_logger(__name__)
+
+
+def smart_resize(
+    height: int, width: int, factor: int = 28, min_pixels: int = 56 * 56, max_pixels: int = 14 * 14 * 4 * 1280
+):
+    """Rescales the image so that the following conditions are met:
+
+    1. Both dimensions (height and width) are divisible by 'factor'.
+
+    2. The total number of pixels is within the range ['min_pixels', 'max_pixels'].
+
+    3. The aspect ratio of the image is maintained as closely as possible.
+
+    """
+    if height < factor or width < factor:
+        raise ValueError(f"height:{height} and width:{width} must be larger than factor:{factor}")
+    elif max(height, width) / min(height, width) > 200:
+        raise ValueError(
+            f"absolute aspect ratio must be smaller than 200, got {max(height, width) / min(height, width)}"
+        )
+    h_bar = round(height / factor) * factor
+    w_bar = round(width / factor) * factor
+    if h_bar * w_bar > max_pixels:
+        beta = math.sqrt((height * width) / max_pixels)
+        h_bar = math.floor(height / beta / factor) * factor
+        w_bar = math.floor(width / beta / factor) * factor
+    elif h_bar * w_bar < min_pixels:
+        beta = math.sqrt(min_pixels / (height * width))
+        h_bar = math.ceil(height * beta / factor) * factor
+        w_bar = math.ceil(width * beta / factor) * factor
+    return h_bar, w_bar
+
+
+class Qwen2VLImageProcessor(BaseImageProcessor):
+    r"""
+    Constructs a Qwen2-VL image processor that dynamically resizes images based on the original images.
+
+    Args:
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to resize the image's (height, width) dimensions.
+        size (`Dict[str, int]`, *optional*, defaults to `{"shortest_edge": 56 * 56, "longest_edge": 28 * 28 * 1280}`):
+            Size of the image after resizing. `shortest_edge` and `longest_edge` keys must be present.
+        resample (`PILImageResampling`, *optional*, defaults to `Resampling.BICUBIC`):
+            Resampling filter to use when resizing the image.
+        do_rescale (`bool`, *optional*, defaults to `True`):
+            Whether to rescale the image by the specified scale `rescale_factor`.
+        rescale_factor (`int` or `float`, *optional*, defaults to `1/255`):
+            Scale factor to use if rescaling the image.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the image.
+        image_mean (`float` or `List[float]`, *optional*, defaults to `[0.48145466, 0.4578275, 0.40821073]`):
+            Mean to use if normalizing the image. This is a float or list of floats for each channel in the image.
+        image_std (`float` or `List[float]`, *optional*, defaults to `[0.26862954, 0.26130258, 0.27577711]`):
+            Standard deviation to use if normalizing the image. This is a float or list of floats for each channel in the image.
+        do_convert_rgb (`bool`, *optional*, defaults to `True`):
+            Whether to convert the image to RGB.
+        min_pixels (`int`, *optional*, defaults to `56 * 56`):
+            The min pixels of the image to resize the image.
+        max_pixels (`int`, *optional*, defaults to `28 * 28 * 1280`):
+            The max pixels of the image to resize the image.
+        patch_size (`int`, *optional*, defaults to 14):
+            The spatial patch size of the vision encoder.
+        temporal_patch_size (`int`, *optional*, defaults to 2):
+            The temporal patch size of the vision encoder.
+        merge_size (`int`, *optional*, defaults to 2):
+            The merge size of the vision encoder to llm encoder.
+    """
+
+    model_input_names = ["pixel_values", "image_grid_thw", "pixel_values_videos", "video_grid_thw"]
+
+    def __init__(
+        self,
+        do_resize: bool = True,
+        size: Optional[Dict[str, int]] = None,
+        resample: PILImageResampling = PILImageResampling.BICUBIC,
+        do_rescale: bool = True,
+        rescale_factor: Union[int, float] = 1 / 255,
+        do_normalize: bool = True,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_convert_rgb: bool = True,
+        min_pixels: Optional[int] = None,
+        max_pixels: Optional[int] = None,
+        patch_size: int = 14,
+        temporal_patch_size: int = 2,
+        merge_size: int = 2,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        if size is not None and ("shortest_edge" not in size or "longest_edge" not in size):
+            raise ValueError("size must contain 'shortest_edge' and 'longest_edge' keys.")
+        else:
+            size = {"shortest_edge": 56 * 56, "longest_edge": 28 * 28 * 1280}
+        # backward compatibility: override size with min_pixels and max_pixels if they are provided
+        if min_pixels is not None:
+            size["shortest_edge"] = min_pixels
+        if max_pixels is not None:
+            size["longest_edge"] = max_pixels
+        self.min_pixels = size["shortest_edge"]
+        self.max_pixels = size["longest_edge"]
+        self.size = size
+
+        self.do_resize = do_resize
+        self.resample = resample
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean if image_mean is not None else OPENAI_CLIP_MEAN
+        self.image_std = image_std if image_std is not None else OPENAI_CLIP_STD
+
+        self.patch_size = patch_size
+        self.temporal_patch_size = temporal_patch_size
+        self.merge_size = merge_size
+        self.do_convert_rgb = do_convert_rgb
+
+    def _preprocess(
+        self,
+        images: Union[ImageInput, VideoInput],
+        do_resize: Optional[bool] = None,
+        size: Optional[Dict[str, int]] = None,
+        resample: PILImageResampling = None,
+        do_rescale: Optional[bool] = None,
+        rescale_factor: Optional[float] = None,
+        do_normalize: Optional[bool] = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        patch_size: Optional[int] = None,
+        temporal_patch_size: Optional[int] = None,
+        merge_size: Optional[int] = None,
+        do_convert_rgb: Optional[bool] = None,
+        data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ):
+        """
+        Preprocess an image or batch of images. Copy of the `preprocess` method from `CLIPImageProcessor`.
+
+        Args:
+            images (`ImageInput`):
+                Image or batch of images to preprocess. Expects pixel values ranging from 0 to 255. If pixel values range from 0 to 1, set `do_rescale=False`.
+            vision_info (`List[Dict]`, *optional*):
+                Optional list of dictionaries containing additional information about vision inputs.
+            do_resize (`bool`, *optional*, defaults to `self.do_resize`):
+                Whether to resize the image.
+            size (`Dict[str, int]`, *optional*, defaults to `self.size`):
+                Size of the image after resizing. `shortest_edge` and `longest_edge` keys must be present.
+            resample (`PILImageResampling`, *optional*, defaults to `self.resample`):
+                Resampling filter to use if resizing the image. This can be one of the `PILImageResampling` enums.
+            do_rescale (`bool`, *optional*, defaults to `self.do_rescale`):
+                Whether to rescale the image.
+            rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`):
+                Scale factor to use if rescaling the image.
+            do_normalize (`bool`, *optional*, defaults to `self.do_normalize`):
+                Whether to normalize the image.
+            image_mean (`float` or `List[float]`, *optional*, defaults to `self.image_mean`):
+                Mean to use if normalizing the image. Can be a float or a list of floats corresponding to the number of channels in the image.
+            image_std (`float` or `List[float]`, *optional*, defaults to `self.image_std`):
+                Standard deviation to use if normalizing the image. Can be a float or a list of floats corresponding to the number of channels in the image.
+            patch_size (`int`, *optional*, defaults to `self.patch_size`):
+                The spatial patch size of the vision encoder.
+            temporal_patch_size (`int`, *optional*, defaults to `self.temporal_patch_size`):
+                The temporal patch size of the vision encoder.
+            merge_size (`int`, *optional*, defaults to `self.merge_size`):
+                The merge size of the vision encoder to llm encoder.
+            do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
+                Whether to convert the image to RGB.
+            data_format (`ChannelDimension`, *optional*, defaults to `ChannelDimension.FIRST`):
+                The channel dimension format for the output image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - Unset: Use the channel dimension format of the input image.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format for the input image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.   - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
+        """
+        images = make_list_of_images(images)
+
+        if do_convert_rgb:
+            images = [convert_to_rgb(image) for image in images]
+
+        # All transformations expect numpy arrays.
+        images = [to_numpy_array(image) for image in images]
+
+        if do_rescale and is_scaled_image(images[0]):
+            logger.warning_once(
+                "It looks like you are trying to rescale already rescaled images. If the input"
+                " images have pixel values between 0 and 1, set `do_rescale=False` to avoid rescaling them again."
+            )
+        if input_data_format is None:
+            # We assume that all images have the same channel dimension format.
+            input_data_format = infer_channel_dimension_format(images[0])
+
+        height, width = get_image_size(images[0], channel_dim=input_data_format)
+        resized_height, resized_width = height, width
+        processed_images = []
+        for image in images:
+            if do_resize:
+                resized_height, resized_width = smart_resize(
+                    height,
+                    width,
+                    factor=patch_size * merge_size,
+                    min_pixels=size["shortest_edge"],
+                    max_pixels=size["longest_edge"],
+                )
+                image = resize(
+                    image, size=(resized_height, resized_width), resample=resample, input_data_format=input_data_format
+                )
+
+            if do_rescale:
+                image = self.rescale(image, scale=rescale_factor, input_data_format=input_data_format)
+
+            if do_normalize:
+                image = self.normalize(
+                    image=image, mean=image_mean, std=image_std, input_data_format=input_data_format
+                )
+
+            image = to_channel_dimension_format(image, data_format, input_channel_dim=input_data_format)
+            processed_images.append(image)
+
+        patches = np.array(processed_images)
+        if data_format == ChannelDimension.LAST:
+            patches = patches.transpose(0, 3, 1, 2)
+        if patches.shape[0] % temporal_patch_size != 0:
+            repeats = np.repeat(
+                patches[-1][np.newaxis], temporal_patch_size - (patches.shape[0] % temporal_patch_size), axis=0
+            )
+            patches = np.concatenate([patches, repeats], axis=0)
+        channel = patches.shape[1]
+        grid_t = patches.shape[0] // temporal_patch_size
+        grid_h, grid_w = resized_height // patch_size, resized_width // patch_size
+        patches = patches.reshape(
+            grid_t,
+            temporal_patch_size,
+            channel,
+            grid_h // merge_size,
+            merge_size,
+            patch_size,
+            grid_w // merge_size,
+            merge_size,
+            patch_size,
+        )
+        patches = patches.transpose(0, 3, 6, 4, 7, 2, 1, 5, 8)
+        flatten_patches = patches.reshape(
+            grid_t * grid_h * grid_w, channel * temporal_patch_size * patch_size * patch_size
+        )
+
+        return flatten_patches, (grid_t, grid_h, grid_w)
+
+    def preprocess(
+        self,
+        images: ImageInput,
+        videos: VideoInput = None,
+        do_resize: Optional[bool] = None,
+        size: Optional[Dict[str, int]] = None,
+        min_pixels: Optional[int] = None,
+        max_pixels: Optional[int] = None,
+        resample: PILImageResampling = None,
+        do_rescale: Optional[bool] = None,
+        rescale_factor: Optional[float] = None,
+        do_normalize: Optional[bool] = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        patch_size: Optional[int] = None,
+        temporal_patch_size: Optional[int] = None,
+        merge_size: Optional[int] = None,
+        do_convert_rgb: Optional[bool] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ):
+        """
+        Args:
+            images (`ImageInput`):
+                Image to preprocess. Expects a single or batch of images with pixel values ranging from 0 to 255. If
+                passing in images with pixel values between 0 and 1, set `do_rescale=False`.
+            videos (`VideoInput`):
+                Video to preprocess. Expects a single or batch of videos with pixel values ranging from 0 to 255. If
+                passing in videos with pixel values between 0 and 1, set `do_rescale=False`.
+            do_resize (`bool`, *optional*, defaults to `self.do_resize`):
+                Whether to resize the image.
+            size (`Dict[str, int]`, *optional*, defaults to `self.size`):
+                Size of the image after resizing. Shortest edge of the image is resized to size["shortest_edge"], with
+                the longest edge resized to keep the input aspect ratio.
+            resample (`int`, *optional*, defaults to `self.resample`):
+                Resampling filter to use if resizing the image. This can be one of the enum `PILImageResampling`. Only
+                has an effect if `do_resize` is set to `True`.
+            do_rescale (`bool`, *optional*, defaults to `self.do_rescale`):
+                Whether to rescale the image.
+            rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`):
+                Rescale factor to rescale the image by if `do_rescale` is set to `True`.
+            do_normalize (`bool`, *optional*, defaults to `self.do_normalize`):
+                Whether to normalize the image.
+            image_mean (`float` or `List[float]`, *optional*, defaults to `self.image_mean`):
+                Image mean to use for normalization. Only has an effect if `do_normalize` is set to `True`.
+            image_std (`float` or `List[float]`, *optional*, defaults to `self.image_std`):
+                Image standard deviation to use for normalization. Only has an effect if `do_normalize` is set to
+                `True`.
+            min_pixels (`int`, *optional*, defaults to `self.min_pixels`):
+                The min pixels of the image to resize the image.
+            max_pixels (`int`, *optional*, defaults to `self.max_pixels`):
+                The max pixels of the image to resize the image.
+            patch_size (`int`, *optional*, defaults to `self.patch_size`):
+                The spatial patch size of the vision encoder.
+            temporal_patch_size (`int`, *optional*, defaults to `self.temporal_patch_size`):
+                The temporal patch size of the vision encoder.
+            merge_size (`int`, *optional*, defaults to `self.merge_size`):
+                The merge size of the vision encoder to llm encoder.
+            do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
+                Whether to convert the image to RGB.
+            return_tensors (`str` or `TensorType`, *optional*):
+                The type of tensors to return. Can be one of:
+                - Unset: Return a list of `np.ndarray`.
+                - `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
+                - `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
+                - `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
+                - `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
+            data_format (`ChannelDimension` or `str`, *optional*, defaults to `ChannelDimension.FIRST`):
+                The channel dimension format for the output image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - Unset: Use the channel dimension format of the input image.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format for the input image. If unset, the channel dimension format is inferred
+                from the input image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
+
+        """
+        min_pixels = min_pixels if min_pixels is not None else self.min_pixels
+        max_pixels = max_pixels if max_pixels is not None else self.max_pixels
+
+        if size is not None:
+            if "shortest_edge" not in size or "longest_edge" not in size:
+                raise ValueError("size must contain 'shortest_edge' and 'longest_edge' keys.")
+            min_pixels = size["shortest_edge"]
+        elif min_pixels is not None and max_pixels is not None:
+            # backward compatibility: override size with min_pixels and max_pixels if they are provided
+            size = {"shortest_edge": min_pixels, "longest_edge": max_pixels}
+        else:
+            size = {**self.size}
+
+        do_resize = do_resize if do_resize is not None else self.do_resize
+
+        resample = resample if resample is not None else self.resample
+        do_rescale = do_rescale if do_rescale is not None else self.do_rescale
+        rescale_factor = rescale_factor if rescale_factor is not None else self.rescale_factor
+        do_normalize = do_normalize if do_normalize is not None else self.do_normalize
+        image_mean = image_mean if image_mean is not None else self.image_mean
+        image_std = image_std if image_std is not None else self.image_std
+        patch_size = patch_size if patch_size is not None else self.patch_size
+        temporal_patch_size = temporal_patch_size if temporal_patch_size is not None else self.temporal_patch_size
+        merge_size = merge_size if merge_size is not None else self.merge_size
+        do_convert_rgb = do_convert_rgb if do_convert_rgb is not None else self.do_convert_rgb
+
+        if images is not None:
+            images = make_flat_list_of_images(images)
+
+        if images is not None and not valid_images(images):
+            raise ValueError(
+                "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
+                "torch.Tensor, tf.Tensor or jax.ndarray."
+            )
+
+        validate_preprocess_arguments(
+            rescale_factor=rescale_factor,
+            do_normalize=do_normalize,
+            image_mean=image_mean,
+            image_std=image_std,
+            do_resize=do_resize,
+            size=size,
+            resample=resample,
+        )
+
+        data = {}
+        if images is not None:
+            pixel_values, vision_grid_thws = [], []
+            for image in images:
+                patches, image_grid_thw = self._preprocess(
+                    image,
+                    do_resize=do_resize,
+                    size=size,
+                    resample=resample,
+                    do_rescale=do_rescale,
+                    rescale_factor=rescale_factor,
+                    do_normalize=do_normalize,
+                    image_mean=image_mean,
+                    image_std=image_std,
+                    patch_size=patch_size,
+                    temporal_patch_size=temporal_patch_size,
+                    merge_size=merge_size,
+                    data_format=data_format,
+                    do_convert_rgb=do_convert_rgb,
+                    input_data_format=input_data_format,
+                )
+                pixel_values.extend(patches)
+                vision_grid_thws.append(image_grid_thw)
+            pixel_values = np.array(pixel_values)
+            vision_grid_thws = np.array(vision_grid_thws)
+            data.update({"pixel_values": pixel_values, "image_grid_thw": vision_grid_thws})
+
+        # kept for BC only and should be removed after v5.0
+        if videos is not None:
+            logger.warning(
+                "`Qwen2VLImageProcessor` works only with image inputs and doesn't process videos anymore. "
+                "This is a deprecated behavior and will be removed in v5.0. "
+                "Your videos should be forwarded to `Qwen2VLVideoProcessor`. "
+            )
+            videos = make_batched_videos(videos)
+            pixel_values_videos, vision_grid_thws_videos = [], []
+            for images in videos:
+                patches, video_grid_thw = self._preprocess(
+                    images,
+                    do_resize=do_resize,
+                    size=size,
+                    resample=resample,
+                    do_rescale=do_rescale,
+                    rescale_factor=rescale_factor,
+                    do_normalize=do_normalize,
+                    image_mean=image_mean,
+                    image_std=image_std,
+                    patch_size=patch_size,
+                    temporal_patch_size=temporal_patch_size,
+                    merge_size=merge_size,
+                    data_format=data_format,
+                    do_convert_rgb=do_convert_rgb,
+                    input_data_format=input_data_format,
+                )
+                pixel_values_videos.extend(patches)
+                vision_grid_thws_videos.append(video_grid_thw)
+            data.update(
+                {
+                    "pixel_values_videos": np.array(pixel_values_videos),
+                    "video_grid_thw": np.array(vision_grid_thws_videos),
+                }
+            )
+
+        return BatchFeature(data=data, tensor_type=return_tensors)
+
+
+__all__ = ["Qwen2VLImageProcessor"]

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2e3f1f331eb3aa68636be825a1b058b147bbf0d6d87af2333d3c317fcb8bdff
+size 1330580320

modeling_qwen2_vl.py

@@ -0,0 +1,410 @@
+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch Qwen2-VL model."""
+
+import math
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import LayerNorm
+
+from transformers.activations import ACT2FN
+from transformers.modeling_flash_attention_utils import is_flash_attn_available
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import auto_docstring, logging
+from .configuration_qwen2_vl import Qwen2VLVisionConfig
+
+
+if is_flash_attn_available():
+    from transformers.modeling_flash_attention_utils import flash_attn_varlen_func
+
+
+
+logger = logging.get_logger(__name__)
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+def apply_rotary_pos_emb_vision(
+    q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    orig_q_dtype = q.dtype
+    orig_k_dtype = k.dtype
+    q, k = q.float(), k.float()
+    cos, sin = cos.unsqueeze(-2).float(), sin.unsqueeze(-2).float()
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    q_embed = q_embed.to(orig_q_dtype)
+    k_embed = k_embed.to(orig_k_dtype)
+    return q_embed, k_embed
+
+
+class VisionRotaryEmbedding(nn.Module):
+    def __init__(self, dim: int, theta: float = 10000.0) -> None:
+        super().__init__()
+        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+    def forward(self, seqlen: int) -> torch.Tensor:
+        seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
+        freqs = torch.outer(seq, self.inv_freq)
+        return freqs
+
+
+class PatchEmbed(nn.Module):
+    def __init__(
+        self,
+        patch_size: int = 14,
+        temporal_patch_size: int = 2,
+        in_channels: int = 3,
+        embed_dim: int = 1152,
+    ) -> None:
+        super().__init__()
+        self.patch_size = patch_size
+        self.temporal_patch_size = temporal_patch_size
+        self.in_channels = in_channels
+        self.embed_dim = embed_dim
+
+        kernel_size = [temporal_patch_size, patch_size, patch_size]
+        self.proj = nn.Conv3d(in_channels, embed_dim, kernel_size=kernel_size, stride=kernel_size, bias=False)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        target_dtype = self.proj.weight.dtype
+        hidden_states = hidden_states.view(
+            -1, self.in_channels, self.temporal_patch_size, self.patch_size, self.patch_size
+        )
+        hidden_states = self.proj(hidden_states.to(dtype=target_dtype)).view(-1, self.embed_dim)
+        return hidden_states
+
+
+class PatchMerger(nn.Module):
+    def __init__(self, dim: int, context_dim: int, spatial_merge_size: int = 2) -> None:
+        super().__init__()
+        self.hidden_size = context_dim * (spatial_merge_size**2)
+        self.ln_q = LayerNorm(context_dim, eps=1e-6)
+        self.mlp = nn.Sequential(
+            nn.Linear(self.hidden_size, self.hidden_size),
+            nn.GELU(),
+            nn.Linear(self.hidden_size, dim),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.mlp(self.ln_q(x).view(-1, self.hidden_size))
+        return x
+
+
+class VisionMlp(nn.Module):
+    def __init__(self, dim: int, hidden_dim: int, hidden_act: str) -> None:
+        super().__init__()
+        self.fc1 = nn.Linear(dim, hidden_dim)
+        self.act = ACT2FN[hidden_act]
+        self.fc2 = nn.Linear(hidden_dim, dim)
+
+    def forward(self, x) -> torch.Tensor:
+        return self.fc2(self.act(self.fc1(x)))
+
+
+class VisionAttention(nn.Module):
+    def __init__(self, dim: int, num_heads: int = 16) -> None:
+        super().__init__()
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.qkv = nn.Linear(dim, dim * 3, bias=True)
+        self.proj = nn.Linear(dim, dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        rotary_pos_emb: Optional[torch.Tensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> torch.Tensor:
+        seq_length = hidden_states.shape[0]
+        q, k, v = self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
+        if position_embeddings is None:
+            logger.warning_once(
+                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+                "through `rotary_pos_emb` (2D tensor of RoPE theta values), to using externally computed "
+                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.54 `rotary_pos_emb` will be "
+                "removed and `position_embeddings` will be mandatory."
+            )
+            emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+        else:
+            cos, sin = position_embeddings
+        q, k = apply_rotary_pos_emb_vision(q, k, cos, sin)
+
+        attention_mask = torch.full(
+            [1, seq_length, seq_length], torch.finfo(q.dtype).min, device=q.device, dtype=q.dtype
+        )
+        for i in range(1, len(cu_seqlens)):
+            attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = 0
+
+        q = q.transpose(0, 1)
+        k = k.transpose(0, 1)
+        v = v.transpose(0, 1)
+        attn_weights = torch.matmul(q, k.transpose(1, 2)) / math.sqrt(self.head_dim)
+        attn_weights = attn_weights + attention_mask
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(q.dtype)
+        attn_output = torch.matmul(attn_weights, v)
+        attn_output = attn_output.transpose(0, 1)
+        attn_output = attn_output.reshape(seq_length, -1)
+        attn_output = self.proj(attn_output)
+        return attn_output
+
+
+class VisionFlashAttention2(nn.Module):
+    def __init__(self, dim: int, num_heads: int = 16) -> None:
+        super().__init__()
+        self.num_heads = num_heads
+        self.qkv = nn.Linear(dim, dim * 3, bias=True)
+        self.proj = nn.Linear(dim, dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        rotary_pos_emb: Optional[torch.Tensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> torch.Tensor:
+        seq_length = hidden_states.shape[0]
+        q, k, v = self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
+        if position_embeddings is None:
+            logger.warning_once(
+                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+                "through `rotary_pos_emb` (2D tensor of RoPE theta values), to using externally computed "
+                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.54 `rotary_pos_emb` will be "
+                "removed and `position_embeddings` will be mandatory."
+            )
+            emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+        else:
+            cos, sin = position_embeddings
+        q, k = apply_rotary_pos_emb_vision(q, k, cos, sin)
+
+        max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
+        attn_output = flash_attn_varlen_func(q, k, v, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen).reshape(
+            seq_length, -1
+        )
+        attn_output = self.proj(attn_output)
+        return attn_output
+
+
+class VisionSdpaAttention(nn.Module):
+    def __init__(self, dim: int, num_heads: int = 16) -> None:
+        super().__init__()
+        self.num_heads = num_heads
+        self.qkv = nn.Linear(dim, dim * 3, bias=True)
+        self.proj = nn.Linear(dim, dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        rotary_pos_emb: Optional[torch.Tensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> torch.Tensor:
+        seq_length = hidden_states.shape[0]
+        q, k, v = self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
+        if position_embeddings is None:
+            logger.warning_once(
+                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+                "through `rotary_pos_emb` (2D tensor of RoPE theta values), to using externally computed "
+                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.54 `rotary_pos_emb` will be "
+                "removed and `position_embeddings` will be mandatory."
+            )
+            emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+        else:
+            cos, sin = position_embeddings
+        q, k = apply_rotary_pos_emb_vision(q, k, cos, sin)
+
+        attention_mask = torch.zeros([1, seq_length, seq_length], device=q.device, dtype=torch.bool)
+        for i in range(1, len(cu_seqlens)):
+            attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = True
+        q = q.transpose(0, 1)
+        k = k.transpose(0, 1)
+        v = v.transpose(0, 1)
+        attn_output = F.scaled_dot_product_attention(
+            q.unsqueeze(0), k.unsqueeze(0), v.unsqueeze(0), attention_mask, dropout_p=0.0
+        )
+        attn_output = attn_output.squeeze(0).transpose(0, 1)
+        attn_output = attn_output.reshape(seq_length, -1)
+        attn_output = self.proj(attn_output)
+        return attn_output
+
+
+QWEN2_VL_VISION_ATTENTION_CLASSES = {
+    "eager": VisionAttention,
+    "flash_attention_2": VisionFlashAttention2,
+    "sdpa": VisionSdpaAttention,
+}
+
+
+class Qwen2VLVisionBlock(nn.Module):
+    def __init__(self, config, attn_implementation: str = "sdpa") -> None:
+        super().__init__()
+        self.norm1 = LayerNorm(config.embed_dim, eps=1e-6)
+        self.norm2 = LayerNorm(config.embed_dim, eps=1e-6)
+        mlp_hidden_dim = int(config.embed_dim * config.mlp_ratio)
+
+        self.attn = QWEN2_VL_VISION_ATTENTION_CLASSES[attn_implementation](
+            config.embed_dim, num_heads=config.num_heads
+        )
+        self.mlp = VisionMlp(dim=config.embed_dim, hidden_dim=mlp_hidden_dim, hidden_act=config.hidden_act)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        rotary_pos_emb: Optional[torch.Tensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> torch.Tensor:
+        hidden_states = hidden_states + self.attn(
+            self.norm1(hidden_states),
+            cu_seqlens=cu_seqlens,
+            rotary_pos_emb=rotary_pos_emb,
+            position_embeddings=position_embeddings,
+        )
+        hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
+        return hidden_states
+
+@auto_docstring
+class Qwen2VisionTransformerPretrainedModel(PreTrainedModel):
+    config_class = Qwen2VLVisionConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Qwen2VLVisionBlock"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = True
+    _supports_static_cache = False  # TODO (joao): fix. torch.compile failing probably due to `cache_positions`
+
+    def __init__(self, config) -> None:
+        super().__init__(config)
+        self.spatial_merge_size = config.spatial_merge_size
+
+        self.patch_embed = PatchEmbed(
+            patch_size=config.patch_size,
+            temporal_patch_size=config.temporal_patch_size,
+            in_channels=config.in_channels,
+            embed_dim=config.embed_dim,
+        )
+
+        head_dim = config.embed_dim // config.num_heads
+        self.rotary_pos_emb = VisionRotaryEmbedding(head_dim // 2)
+
+        self.blocks = nn.ModuleList(
+            [Qwen2VLVisionBlock(config, config._attn_implementation) for _ in range(config.depth)]
+        )
+        self.merger = PatchMerger(
+            dim=config.hidden_size, context_dim=config.embed_dim, spatial_merge_size=config.spatial_merge_size
+        )
+        self.gradient_checkpointing = False
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, (nn.Linear, nn.Conv3d)):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    def get_dtype(self) -> torch.dtype:
+        return self.blocks[0].mlp.fc2.weight.dtype
+
+    def get_device(self) -> torch.device:
+        return self.blocks[0].mlp.fc2.weight.device
+
+    def rot_pos_emb(self, grid_thw):
+        pos_ids = []
+        for t, h, w in grid_thw:
+            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
+            hpos_ids = hpos_ids.reshape(
+                h // self.spatial_merge_size,
+                self.spatial_merge_size,
+                w // self.spatial_merge_size,
+                self.spatial_merge_size,
+            )
+            hpos_ids = hpos_ids.permute(0, 2, 1, 3)
+            hpos_ids = hpos_ids.flatten()
+
+            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
+            wpos_ids = wpos_ids.reshape(
+                h // self.spatial_merge_size,
+                self.spatial_merge_size,
+                w // self.spatial_merge_size,
+                self.spatial_merge_size,
+            )
+            wpos_ids = wpos_ids.permute(0, 2, 1, 3)
+            wpos_ids = wpos_ids.flatten()
+            pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
+        pos_ids = torch.cat(pos_ids, dim=0)
+        max_grid_size = grid_thw[:, 1:].max()
+        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
+        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
+        return rotary_pos_emb
+
+    @auto_docstring
+    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor) -> torch.Tensor:
+        r"""
+        grid_thw (`torch.LongTensor` of shape `(num_images, 3)`):
+            The temporal, height and width dimensions of feature shape for each image. Each row contains [t, h, w] values.
+        """
+        hidden_states = self.patch_embed(hidden_states)
+        rotary_pos_emb = self.rot_pos_emb(grid_thw)
+        emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
+        position_embeddings = (emb.cos(), emb.sin())
+
+        cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]).cumsum(
+            dim=0,
+            # Select dtype based on the following factors:
+            #  - FA2 requires that cu_seqlens_q must have dtype int32
+            #  - torch.onnx.export requires that cu_seqlens_q must have same dtype as grid_thw
+            # See https://github.com/huggingface/transformers/pull/34852 for more information
+            dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
+        )
+        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
+
+        for blk in self.blocks:
+            if self.gradient_checkpointing and self.training:
+                hidden_states = self._gradient_checkpointing_func(
+                    blk.__call__, hidden_states, cu_seqlens, None, position_embeddings
+                )
+            else:
+                hidden_states = blk(hidden_states, cu_seqlens=cu_seqlens, position_embeddings=position_embeddings)
+
+        return self.merger(hidden_states)
+

preprocessor_config.json

@@ -0,0 +1,30 @@
+{
+  "auto_map": {
+    "AutoImageProcessor": "image_processing_qwen2_vl.Qwen2VLImageProcessor"
+  },
+  "do_convert_rgb": true,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.48145466,
+    0.4578275,
+    0.40821073
+  ],
+  "image_std": [
+    0.26862954,
+    0.26130258,
+    0.27577711
+  ],
+  "max_pixels": 12845056,
+  "merge_size": 2,
+  "min_pixels": 3136,
+  "patch_size": 14,
+  "resample": 3,
+  "rescale_factor": 0.00392156862745098,
+  "size": {
+    "longest_edge": 12845056,
+    "shortest_edge": 3136
+  },
+  "temporal_patch_size": 2
+}